1 //===- InlineFunction.cpp - Code to perform function inlining -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements inlining of a function into a call site, resolving
11 // parameters and the return value as appropriate.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Utils/Cloning.h"
16 #include "llvm/Constants.h"
17 #include "llvm/DerivedTypes.h"
18 #include "llvm/Module.h"
19 #include "llvm/Instructions.h"
20 #include "llvm/Intrinsics.h"
21 #include "llvm/ParameterAttributes.h"
22 #include "llvm/Analysis/CallGraph.h"
23 #include "llvm/Target/TargetData.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/Support/CallSite.h"
28 bool llvm::InlineFunction(CallInst *CI, CallGraph *CG, const TargetData *TD) {
29 return InlineFunction(CallSite(CI), CG, TD);
31 bool llvm::InlineFunction(InvokeInst *II, CallGraph *CG, const TargetData *TD) {
32 return InlineFunction(CallSite(II), CG, TD);
35 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
36 /// in the body of the inlined function into invokes and turn unwind
37 /// instructions into branches to the invoke unwind dest.
39 /// II is the invoke instruction begin inlined. FirstNewBlock is the first
40 /// block of the inlined code (the last block is the end of the function),
41 /// and InlineCodeInfo is information about the code that got inlined.
42 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
43 ClonedCodeInfo &InlinedCodeInfo) {
44 BasicBlock *InvokeDest = II->getUnwindDest();
45 std::vector<Value*> InvokeDestPHIValues;
47 // If there are PHI nodes in the unwind destination block, we need to
48 // keep track of which values came into them from this invoke, then remove
49 // the entry for this block.
50 BasicBlock *InvokeBlock = II->getParent();
51 for (BasicBlock::iterator I = InvokeDest->begin(); isa<PHINode>(I); ++I) {
52 PHINode *PN = cast<PHINode>(I);
53 // Save the value to use for this edge.
54 InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(InvokeBlock));
57 Function *Caller = FirstNewBlock->getParent();
59 // The inlined code is currently at the end of the function, scan from the
60 // start of the inlined code to its end, checking for stuff we need to
62 if (InlinedCodeInfo.ContainsCalls || InlinedCodeInfo.ContainsUnwinds) {
63 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
65 if (InlinedCodeInfo.ContainsCalls) {
66 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ){
67 Instruction *I = BBI++;
69 // We only need to check for function calls: inlined invoke
70 // instructions require no special handling.
71 if (!isa<CallInst>(I)) continue;
72 CallInst *CI = cast<CallInst>(I);
74 // If this call cannot unwind, don't convert it to an invoke.
75 if (CI->doesNotThrow())
78 // Convert this function call into an invoke instruction.
79 // First, split the basic block.
80 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
82 // Next, create the new invoke instruction, inserting it at the end
83 // of the old basic block.
84 SmallVector<Value*, 8> InvokeArgs(CI->op_begin()+1, CI->op_end());
86 new InvokeInst(CI->getCalledValue(), Split, InvokeDest,
87 InvokeArgs.begin(), InvokeArgs.end(),
88 CI->getName(), BB->getTerminator());
89 II->setCallingConv(CI->getCallingConv());
90 II->setParamAttrs(CI->getParamAttrs());
92 // Make sure that anything using the call now uses the invoke!
93 CI->replaceAllUsesWith(II);
95 // Delete the unconditional branch inserted by splitBasicBlock
96 BB->getInstList().pop_back();
97 Split->getInstList().pop_front(); // Delete the original call
99 // Update any PHI nodes in the exceptional block to indicate that
100 // there is now a new entry in them.
102 for (BasicBlock::iterator I = InvokeDest->begin();
103 isa<PHINode>(I); ++I, ++i) {
104 PHINode *PN = cast<PHINode>(I);
105 PN->addIncoming(InvokeDestPHIValues[i], BB);
108 // This basic block is now complete, start scanning the next one.
113 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
114 // An UnwindInst requires special handling when it gets inlined into an
115 // invoke site. Once this happens, we know that the unwind would cause
116 // a control transfer to the invoke exception destination, so we can
117 // transform it into a direct branch to the exception destination.
118 new BranchInst(InvokeDest, UI);
120 // Delete the unwind instruction!
121 UI->getParent()->getInstList().pop_back();
123 // Update any PHI nodes in the exceptional block to indicate that
124 // there is now a new entry in them.
126 for (BasicBlock::iterator I = InvokeDest->begin();
127 isa<PHINode>(I); ++I, ++i) {
128 PHINode *PN = cast<PHINode>(I);
129 PN->addIncoming(InvokeDestPHIValues[i], BB);
135 // Now that everything is happy, we have one final detail. The PHI nodes in
136 // the exception destination block still have entries due to the original
137 // invoke instruction. Eliminate these entries (which might even delete the
139 InvokeDest->removePredecessor(II->getParent());
142 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
143 /// into the caller, update the specified callgraph to reflect the changes we
144 /// made. Note that it's possible that not all code was copied over, so only
145 /// some edges of the callgraph will be remain.
146 static void UpdateCallGraphAfterInlining(const Function *Caller,
147 const Function *Callee,
148 Function::iterator FirstNewBlock,
149 DenseMap<const Value*, Value*> &ValueMap,
151 // Update the call graph by deleting the edge from Callee to Caller
152 CallGraphNode *CalleeNode = CG[Callee];
153 CallGraphNode *CallerNode = CG[Caller];
154 CallerNode->removeCallEdgeTo(CalleeNode);
156 // Since we inlined some uninlined call sites in the callee into the caller,
157 // add edges from the caller to all of the callees of the callee.
158 for (CallGraphNode::iterator I = CalleeNode->begin(),
159 E = CalleeNode->end(); I != E; ++I) {
160 const Instruction *OrigCall = I->first.getInstruction();
162 DenseMap<const Value*, Value*>::iterator VMI = ValueMap.find(OrigCall);
163 // Only copy the edge if the call was inlined!
164 if (VMI != ValueMap.end() && VMI->second) {
165 // If the call was inlined, but then constant folded, there is no edge to
166 // add. Check for this case.
167 if (Instruction *NewCall = dyn_cast<Instruction>(VMI->second))
168 CallerNode->addCalledFunction(CallSite::get(NewCall), I->second);
174 // InlineFunction - This function inlines the called function into the basic
175 // block of the caller. This returns false if it is not possible to inline this
176 // call. The program is still in a well defined state if this occurs though.
178 // Note that this only does one level of inlining. For example, if the
179 // instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
180 // exists in the instruction stream. Similiarly this will inline a recursive
181 // function by one level.
183 bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) {
184 Instruction *TheCall = CS.getInstruction();
185 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
186 "Instruction not in function!");
188 const Function *CalledFunc = CS.getCalledFunction();
189 if (CalledFunc == 0 || // Can't inline external function or indirect
190 CalledFunc->isDeclaration() || // call, or call to a vararg function!
191 CalledFunc->getFunctionType()->isVarArg()) return false;
194 // If the call to the callee is a non-tail call, we must clear the 'tail'
195 // flags on any calls that we inline.
196 bool MustClearTailCallFlags =
197 isa<CallInst>(TheCall) && !cast<CallInst>(TheCall)->isTailCall();
199 // If the call to the callee cannot throw, set the 'nounwind' flag on any
200 // calls that we inline.
201 bool MarkNoUnwind = CS.doesNotThrow();
203 BasicBlock *OrigBB = TheCall->getParent();
204 Function *Caller = OrigBB->getParent();
206 // GC poses two hazards to inlining, which only occur when the callee has GC:
207 // 1. If the caller has no GC, then the callee's GC must be propagated to the
209 // 2. If the caller has a differing GC, it is invalid to inline.
210 if (CalledFunc->hasCollector()) {
211 if (!Caller->hasCollector())
212 Caller->setCollector(CalledFunc->getCollector());
213 else if (CalledFunc->getCollector() != Caller->getCollector())
217 // Get an iterator to the last basic block in the function, which will have
218 // the new function inlined after it.
220 Function::iterator LastBlock = &Caller->back();
222 // Make sure to capture all of the return instructions from the cloned
224 std::vector<ReturnInst*> Returns;
225 ClonedCodeInfo InlinedFunctionInfo;
226 Function::iterator FirstNewBlock;
228 { // Scope to destroy ValueMap after cloning.
229 DenseMap<const Value*, Value*> ValueMap;
231 assert(std::distance(CalledFunc->arg_begin(), CalledFunc->arg_end()) ==
232 std::distance(CS.arg_begin(), CS.arg_end()) &&
233 "No varargs calls can be inlined!");
235 // Calculate the vector of arguments to pass into the function cloner, which
236 // matches up the formal to the actual argument values.
237 CallSite::arg_iterator AI = CS.arg_begin();
239 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
240 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
241 Value *ActualArg = *AI;
243 // When byval arguments actually inlined, we need to make the copy implied
244 // by them explicit. However, we don't do this if the callee is readonly
245 // or readnone, because the copy would be unneeded: the callee doesn't
246 // modify the struct.
247 if (CalledFunc->paramHasAttr(ArgNo+1, ParamAttr::ByVal) &&
248 !CalledFunc->onlyReadsMemory()) {
249 const Type *AggTy = cast<PointerType>(I->getType())->getElementType();
250 const Type *VoidPtrTy = PointerType::getUnqual(Type::Int8Ty);
252 // Create the alloca. If we have TargetData, use nice alignment.
254 if (TD) Align = TD->getPrefTypeAlignment(AggTy);
255 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, I->getName(),
256 Caller->begin()->begin());
258 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(),
259 Intrinsic::memcpy_i64);
260 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall);
261 Value *SrcCast = new BitCastInst(*AI, VoidPtrTy, "tmp", TheCall);
265 Size = ConstantExpr::getSizeOf(AggTy);
267 Size = ConstantInt::get(Type::Int64Ty, TD->getTypeStoreSize(AggTy));
269 // Always generate a memcpy of alignment 1 here because we don't know
270 // the alignment of the src pointer. Other optimizations can infer
272 Value *CallArgs[] = {
273 DestCast, SrcCast, Size, ConstantInt::get(Type::Int32Ty, 1)
275 CallInst *TheMemCpy =
276 new CallInst(MemCpyFn, CallArgs, CallArgs+4, "", TheCall);
278 // If we have a call graph, update it.
280 CallGraphNode *MemCpyCGN = CG->getOrInsertFunction(MemCpyFn);
281 CallGraphNode *CallerNode = (*CG)[Caller];
282 CallerNode->addCalledFunction(TheMemCpy, MemCpyCGN);
285 // Uses of the argument in the function should use our new alloca
287 ActualArg = NewAlloca;
290 ValueMap[I] = ActualArg;
293 // We want the inliner to prune the code as it copies. We would LOVE to
294 // have no dead or constant instructions leftover after inlining occurs
295 // (which can happen, e.g., because an argument was constant), but we'll be
296 // happy with whatever the cloner can do.
297 CloneAndPruneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i",
298 &InlinedFunctionInfo, TD);
300 // Remember the first block that is newly cloned over.
301 FirstNewBlock = LastBlock; ++FirstNewBlock;
303 // Update the callgraph if requested.
305 UpdateCallGraphAfterInlining(Caller, CalledFunc, FirstNewBlock, ValueMap,
309 // If there are any alloca instructions in the block that used to be the entry
310 // block for the callee, move them to the entry block of the caller. First
311 // calculate which instruction they should be inserted before. We insert the
312 // instructions at the end of the current alloca list.
315 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
316 for (BasicBlock::iterator I = FirstNewBlock->begin(),
317 E = FirstNewBlock->end(); I != E; )
318 if (AllocaInst *AI = dyn_cast<AllocaInst>(I++)) {
319 // If the alloca is now dead, remove it. This often occurs due to code
321 if (AI->use_empty()) {
322 AI->eraseFromParent();
326 if (isa<Constant>(AI->getArraySize())) {
327 // Scan for the block of allocas that we can move over, and move them
329 while (isa<AllocaInst>(I) &&
330 isa<Constant>(cast<AllocaInst>(I)->getArraySize()))
333 // Transfer all of the allocas over in a block. Using splice means
334 // that the instructions aren't removed from the symbol table, then
336 Caller->getEntryBlock().getInstList().splice(
338 FirstNewBlock->getInstList(),
344 // If the inlined code contained dynamic alloca instructions, wrap the inlined
345 // code with llvm.stacksave/llvm.stackrestore intrinsics.
346 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
347 Module *M = Caller->getParent();
348 const Type *BytePtr = PointerType::getUnqual(Type::Int8Ty);
349 // Get the two intrinsics we care about.
350 Constant *StackSave, *StackRestore;
351 StackSave = M->getOrInsertFunction("llvm.stacksave", BytePtr, NULL);
352 StackRestore = M->getOrInsertFunction("llvm.stackrestore", Type::VoidTy,
355 // If we are preserving the callgraph, add edges to the stacksave/restore
356 // functions for the calls we insert.
357 CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0;
359 // We know that StackSave/StackRestore are Function*'s, because they are
360 // intrinsics which must have the right types.
361 StackSaveCGN = CG->getOrInsertFunction(cast<Function>(StackSave));
362 StackRestoreCGN = CG->getOrInsertFunction(cast<Function>(StackRestore));
363 CallerNode = (*CG)[Caller];
366 // Insert the llvm.stacksave.
367 CallInst *SavedPtr = new CallInst(StackSave, "savedstack",
368 FirstNewBlock->begin());
369 if (CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN);
371 // Insert a call to llvm.stackrestore before any return instructions in the
373 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
374 CallInst *CI = new CallInst(StackRestore, SavedPtr, "", Returns[i]);
375 if (CG) CallerNode->addCalledFunction(CI, StackRestoreCGN);
378 // Count the number of StackRestore calls we insert.
379 unsigned NumStackRestores = Returns.size();
381 // If we are inlining an invoke instruction, insert restores before each
382 // unwind. These unwinds will be rewritten into branches later.
383 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) {
384 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
386 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
387 new CallInst(StackRestore, SavedPtr, "", UI);
393 // If we are inlining tail call instruction through a call site that isn't
394 // marked 'tail', we must remove the tail marker for any calls in the inlined
395 // code. Also, calls inlined through a 'nounwind' call site should be marked
397 if (InlinedFunctionInfo.ContainsCalls &&
398 (MustClearTailCallFlags || MarkNoUnwind)) {
399 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
401 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
402 if (CallInst *CI = dyn_cast<CallInst>(I)) {
403 if (MustClearTailCallFlags)
404 CI->setTailCall(false);
406 CI->setDoesNotThrow();
410 // If we are inlining through a 'nounwind' call site then any inlined 'unwind'
411 // instructions are unreachable.
412 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind)
413 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
415 TerminatorInst *Term = BB->getTerminator();
416 if (isa<UnwindInst>(Term)) {
417 new UnreachableInst(Term);
418 BB->getInstList().erase(Term);
422 // If we are inlining for an invoke instruction, we must make sure to rewrite
423 // any inlined 'unwind' instructions into branches to the invoke exception
424 // destination, and call instructions into invoke instructions.
425 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
426 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
428 // If we cloned in _exactly one_ basic block, and if that block ends in a
429 // return instruction, we splice the body of the inlined callee directly into
430 // the calling basic block.
431 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
432 // Move all of the instructions right before the call.
433 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
434 FirstNewBlock->begin(), FirstNewBlock->end());
435 // Remove the cloned basic block.
436 Caller->getBasicBlockList().pop_back();
438 // If the call site was an invoke instruction, add a branch to the normal
440 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
441 new BranchInst(II->getNormalDest(), TheCall);
443 // If the return instruction returned a value, replace uses of the call with
444 // uses of the returned value.
445 if (!TheCall->use_empty()) {
446 ReturnInst *R = Returns[0];
447 if (R->getNumOperands() > 1) {
448 // Multiple return values.
449 while (!TheCall->use_empty()) {
450 GetResultInst *GR = cast<GetResultInst>(TheCall->use_back());
451 Value *RV = R->getOperand(GR->getIndex());
452 GR->replaceAllUsesWith(RV);
453 GR->eraseFromParent();
456 TheCall->replaceAllUsesWith(R->getReturnValue());
458 // Since we are now done with the Call/Invoke, we can delete it.
459 TheCall->getParent()->getInstList().erase(TheCall);
461 // Since we are now done with the return instruction, delete it also.
462 Returns[0]->getParent()->getInstList().erase(Returns[0]);
464 // We are now done with the inlining.
468 // Otherwise, we have the normal case, of more than one block to inline or
469 // multiple return sites.
471 // We want to clone the entire callee function into the hole between the
472 // "starter" and "ender" blocks. How we accomplish this depends on whether
473 // this is an invoke instruction or a call instruction.
474 BasicBlock *AfterCallBB;
475 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
477 // Add an unconditional branch to make this look like the CallInst case...
478 BranchInst *NewBr = new BranchInst(II->getNormalDest(), TheCall);
480 // Split the basic block. This guarantees that no PHI nodes will have to be
481 // updated due to new incoming edges, and make the invoke case more
482 // symmetric to the call case.
483 AfterCallBB = OrigBB->splitBasicBlock(NewBr,
484 CalledFunc->getName()+".exit");
486 } else { // It's a call
487 // If this is a call instruction, we need to split the basic block that
488 // the call lives in.
490 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
491 CalledFunc->getName()+".exit");
494 // Change the branch that used to go to AfterCallBB to branch to the first
495 // basic block of the inlined function.
497 TerminatorInst *Br = OrigBB->getTerminator();
498 assert(Br && Br->getOpcode() == Instruction::Br &&
499 "splitBasicBlock broken!");
500 Br->setOperand(0, FirstNewBlock);
503 // Now that the function is correct, make it a little bit nicer. In
504 // particular, move the basic blocks inserted from the end of the function
505 // into the space made by splitting the source basic block.
507 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
508 FirstNewBlock, Caller->end());
510 // Handle all of the return instructions that we just cloned in, and eliminate
511 // any users of the original call/invoke instruction.
512 if (Returns.size() > 1) {
513 // The PHI node should go at the front of the new basic block to merge all
514 // possible incoming values.
517 if (!TheCall->use_empty()) {
518 PHI = new PHINode(CalledFunc->getReturnType(),
519 TheCall->getName(), AfterCallBB->begin());
521 // Anything that used the result of the function call should now use the
522 // PHI node as their operand.
524 TheCall->replaceAllUsesWith(PHI);
527 // Loop over all of the return instructions, turning them into unconditional
528 // branches to the merge point now, and adding entries to the PHI node as
530 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
531 ReturnInst *RI = Returns[i];
534 assert(RI->getReturnValue() && "Ret should have value!");
535 assert(RI->getReturnValue()->getType() == PHI->getType() &&
536 "Ret value not consistent in function!");
537 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
540 // Add a branch to the merge point where the PHI node lives if it exists.
541 new BranchInst(AfterCallBB, RI);
543 // Delete the return instruction now
544 RI->getParent()->getInstList().erase(RI);
547 } else if (!Returns.empty()) {
548 // Otherwise, if there is exactly one return value, just replace anything
549 // using the return value of the call with the computed value.
550 if (!TheCall->use_empty())
551 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
553 // Splice the code from the return block into the block that it will return
554 // to, which contains the code that was after the call.
555 BasicBlock *ReturnBB = Returns[0]->getParent();
556 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
557 ReturnBB->getInstList());
559 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
560 ReturnBB->replaceAllUsesWith(AfterCallBB);
562 // Delete the return instruction now and empty ReturnBB now.
563 Returns[0]->eraseFromParent();
564 ReturnBB->eraseFromParent();
565 } else if (!TheCall->use_empty()) {
566 // No returns, but something is using the return value of the call. Just
568 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
571 // Since we are now done with the Call/Invoke, we can delete it.
572 TheCall->eraseFromParent();
574 // We should always be able to fold the entry block of the function into the
575 // single predecessor of the block...
576 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
577 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
579 // Splice the code entry block into calling block, right before the
580 // unconditional branch.
581 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
582 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
584 // Remove the unconditional branch.
585 OrigBB->getInstList().erase(Br);
587 // Now we can remove the CalleeEntry block, which is now empty.
588 Caller->getBasicBlockList().erase(CalleeEntry);